Very Computer

>     I am quite puzzled with the rather poor performance of our raid
> array. The array consists of three 15k U160 hard drives (Seagate
> ST318452LW) on a single SCSI bus, configured as a RAID-0 array
> (striping). The SCSI RAID controller is an Adaptec SCSI RAID 2110S on a
> 64bit-33MHz PCI slot (Tyan AMD760-MP (not MPX) based motherboard). The
> main problem is tha rather low transfer rate of the array (80-85 MB/s
> reading or writing), no matter whether the array be configured as a
> hardware RAID array or a software one (what follows is therefore
> applicable to both types of arrays. In either case, the optimal chunk
> size turns out to be 128kB.

>     For testing, we wrote a C program that dumps large files (2GB) in
> 256MB chunks. The program uses about 256MB of memory out of 512MB
> available, so there is not much room left for disk caching (roughly
> 130MB). Here are performance data.

> 1) Controler configuration

> 1.a SMARTROM : enable extended Int13, 16MB WB or WT cache (does not make
> any difference)

> 1.b CONTROLER : reports a host bus transfer rate of 528MB/s while this
> should be 264 MB/sec (33MHz-64bits PCI instead of 66MHz). Maybee this is
> only a figure aimed at showing off the controller's peak performance

> 1.c BUS : Ultra160, 16 bits, auto termination, 160MB/sec. Bus boot enabled.

> 1.d DRIVES : enabled fields are wide16, synchronous, SMART, Cmd queuing,
> linked cmds, SCSI-3. 160MB/s max transfer rate, optimal status

> 2) OS configuration

> RH Linux 7.3, kernel 2.4.18-3smp, dpt_i2o driver, ext3 filesystem

> 3) Performance data (sustained transfer rates)

> 3.a No RAID array

>    - Disks tested individually : 45MB/s writing, 50 MB/s reading each disk
>    - 2 disks tested simultaneously : 30MB/sec writing, 50 MB/sec reading
> per disk
>    - 3 disks tested simultaneously : 20MB/sec writing, 30 MB/sec reading
> per disk

> 3.b RAID-0 array (3 disks, soft or hard RAID, 128kB chunk size

>    - 80-85 MB/s writing or reading

> As you can see, we are quite far from the 120MB/s (3/4 of 160 MB/s) one
> can expect from a RAID-0 array made of 3 U160 disks, each capable of
> sustaining 50MB/sec, hosted by a 64bit-33MHz PCI controler. As the
> single disk tests show (3.a), we suspect there is a problem on the U160
> SCSI bus itself, which seems to be completely overloaded.

> Finally, we found the controller to be quite unstable because of heat
> problems (even in a well-cooled case). The system indeed freezes upon
> extensive use of the SCSI disks (which by the way are in a separate
> case) and controller unless the PC case is left open.

> How can we help improve the performance of this RAID array ?

> Best Regards,

>     Xavier

>     I am quite puzzled with the rather poor performance of our raid
> array. The array consists of three 15k U160 hard drives (Seagate
> ST318452LW) on a single SCSI bus, configured as a RAID-0 array
> (striping). The SCSI RAID controller is an Adaptec SCSI RAID 2110S on a
> 64bit-33MHz PCI slot (Tyan AMD760-MP (not MPX) based motherboard). The
> main problem is tha rather low transfer rate of the array (80-85 MB/s
> reading or writing), no matter whether the array be configured as a
> hardware RAID array or a software one (what follows is therefore
> applicable to both types of arrays. In either case, the optimal chunk
> size turns out to be 128kB.

>     For testing, we wrote a C program that dumps large files (2GB) in
> 256MB chunks. The program uses about 256MB of memory out of 512MB
> available, so there is not much room left for disk caching (roughly
> 130MB). Here are performance data.

> 1) Controler configuration

> 1.a SMARTROM : enable extended Int13, 16MB WB or WT cache (does not make
> any difference)

> 1.b CONTROLER : reports a host bus transfer rate of 528MB/s while this
> should be 264 MB/sec (33MHz-64bits PCI instead of 66MHz). Maybee this is
> only a figure aimed at showing off the controller's peak performance

> 1.c BUS : Ultra160, 16 bits, auto termination, 160MB/sec. Bus boot enabled.

> 1.d DRIVES : enabled fields are wide16, synchronous, SMART, Cmd queuing,
> linked cmds, SCSI-3. 160MB/s max transfer rate, optimal status

> 2) OS configuration

> RH Linux 7.3, kernel 2.4.18-3smp, dpt_i2o driver, ext3 filesystem

> 3) Performance data (sustained transfer rates)

> 3.a No RAID array

>    - Disks tested individually : 45MB/s writing, 50 MB/s reading each disk
>    - 2 disks tested simultaneously : 30MB/sec writing, 50 MB/sec reading
> per disk
>    - 3 disks tested simultaneously : 20MB/sec writing, 30 MB/sec reading
> per disk

> 3.b RAID-0 array (3 disks, soft or hard RAID, 128kB chunk size

>    - 80-85 MB/s writing or reading

> As you can see, we are quite far from the 120MB/s (3/4 of 160 MB/s) one
> can expect from a RAID-0 array made of 3 U160 disks, each capable of
> sustaining 50MB/sec, hosted by a 64bit-33MHz PCI controler. As the
> single disk tests show (3.a), we suspect there is a problem on the U160
> SCSI bus itself, which seems to be completely overloaded.

> Finally, we found the controller to be quite unstable because of heat
> problems (even in a well-cooled case). The system indeed freezes upon
> extensive use of the SCSI disks (which by the way are in a separate
> case) and controller unless the PC case is left open.

> How can we help improve the performance of this RAID array ?